alexenge/hu-neuro-pipeline

Single trial EEG pipeline at the Abdel Rahman Lab for Neurocognitive Psychology, Humboldt-Universität zu Berlin

Welcome 👋

I’m Alex and I’m currently a PhD student in Psychology.

• 🧑🏻‍💻 R & Python
• 🔬 Semantic processing & visual perception
• 🧠 EEG & fMRI data
• 🏃🏻 Distance runner

hu-neuro-pipeline's Issues

Add command line interface + Docker image

Rework how evokeds are computed

Epochs.average() in MNE 0.24.0 supports creating separate averages for each event_id (which we'd need to create from the metadata first); see https://mne.tools/stable/generated/mne.Epochs.html#mne.Epochs.average.

Always return `time` in seconds, not milliseconds

Make output of `reject_peak_to_peak` more consistent

Add bad_channels to trials

Create automatic release notes from future pull request

Add cluster depth tests and/or ANOVA/LMM-based CBPTs

• For cluster depth tests, see Frossard & Renaud (2022) and https://github.com/jaromilfrossard/mne-python
• For ANOVA-based cluster-based CBPTs, see https://jaromilfrossard.github.io/permuco4brain/

Only published GitHub releases should trigger action

Somehow, each release currently triggers 3 actions, leading 2 of them to fail (because file exists on PyPI).

See https://docs.github.com/en/actions/using-workflows/events-that-trigger-workflows#release

Saving config fails when `log_files` are data frames

Because they can't be serialized to JSON

Detect channels based on standard errors

See https://mne.tools/stable/generated/mne.Epochs.html?highlight=standard_error#mne.Epochs.standard_error and https://doi.org/10.1111/psyp.13793

Improve participant-specific returns in `config`

• Empty auto_bad_channels should be [], not None

• Return indices excluded ICA components per participant

• Return dict with participant IDs (instead of list) for auto_bad_channels, rejected_epochs, excl_ica_components

Adjust default frequencies and cycles for TFR

Improve default options

• Get rid of reject_flat for simplicity
• Increase lowest TFR frequency to 5 Hz to avoid edge effects and have a sufficiently long baseline (see, e.g., recommendations by Roach & Mathalon, 2008)
• Adjust wavelet number of cycles accordingly and align better with EEGLAB (3 to 10 cycles in steps of 0.5)

Add option to remove evoked activity from TFR

Make input arguments more consistent

• baseline should become baseline_tmin and baseline_tmax

• Use auto instead of ICA method (e.g., 'fastica') for ocular_correction

• Update docs/ accordingly

Add tests

Based on publicly available data from our lab as in the original MATLAB pipeline.

Add boilerplate

pipeline() should print some paragraphs for the methods section as is done, e.g., in NiMARE.

Add workshop and UCAP example repos as submodules

R examples don't work for single components

Because R converts c("some_string") to "some_string, which makes pipeline add an additional layer of lists, which produces an error when roi was already specified as an (actual) list. Sigh.

Proper documentation

Using Read the Docs or the like

Bad channels don't actually get interpolated

Stupid me 😓

Automatically detect missing EEG trials

Fix automatic GitHub release names

Export pipeline config to JSON

Allow specifying the number of extracted components for ICA

Add different input file types

Interpolate bad channels using `pyprep`

See https://pyprep.readthedocs.io/en/latest/generated/pyprep.PrepPipeline.html#pyprep.PrepPipeline

Add permutation tests

Compute tempo-spatial one sample tests for each combination of condition_cols.

Improve UCAP fetcher to get an arbitrary number of participants

Return pipeline `config` as dict and save to JSON

This enables to:

• Keep track of (automatically detected) bad channels per participant
• Keep track of rejected epochs per participant
• Re-use input arguments as inline code in the methods section

Provide functionality to automatically combine raw data sets

Sometimes one participant has 2 or more raw datasets because the recording was stopped accidently

Match epochs also in the case of missing *log file* trials

Currently, we only deal with missing EEG epochs, not with missing log file trials. This seems to happen extremely rarely.

Add BIDS interface

See MNE-BIDS and MNE-BIDS-Pipeline

Prettify logging

Get rid of the many info messages created, e.g., for the HTML reports

Fail when channels in `components` or `components_tfr` are not in the data

Currently, the pipeline doesn't seem to fail when one or more channels provided via components_tfr['roi'] doesn't exist in the data

If `tfr_subtract_evoked`, TFR gets computed on filtered rather than unfiltered epochs

Missing EEG epochs should go into the config

Specifying triggers_column automatically gets rid of trials which are in the log file but not in the EEG, but right now it only prints this information on screen. The missing trial indices should also go into a separate field in the config (similar to rejected_epochs).

Current single trial pre-stimulus baseline for TFR leads to biased results

Problem: Single trial divisive baseline correction using the pre-stimulus interval leads to a positive bias in post-stimulus power (Grandchamp & Delorme, 2011; Hu et al., 2014)

Solutions:

• Subtractive baseline (Hu et al., 2014) -> But: Doesn't get rid of $1/f$
• Divisive baseline using the full epoch (Grandchamp & Delorme, 2011) (+ additional subtractive baseline using the pre-stimulus interval)

Simplify return of evokeds

pipeline() should return a single DataFrame of evokes instead of a dict. Makes it more predictable in the case of one vs. multiple condition_cols and easier to use the outputs in R.

Add MNE reports for QC

See MNE docs

Make reports more informative

• No. of rejected epochs
• Report / plot single trial mean ERPs per condition?
• ... ?

Compute mixed models with `pymer4` and/or `bambi`

• Pass model formula as a string (new input option) or as column names (reworking average_by)?
• How to apply contrasts and follow-up tests?
• See pymer4
• Also check out statsmodels and MixedModels.jl + pyjulia as alternatives to lme4
• Or use bambi for Bayesian (G)LMMs

Naturalsort files if directories are provided

To deal with (log) files that don't have leading zeros in their filenames, while the EEG and cali files do

`.fif` files for evokeds are not saved with `overwrite=True`

Whereas we do overwrite all other outputs

Make it possible to skip baseline correction

For both ERPs and TFR we shouldn't do baseline correction if both baseline_tmin and baseline_tmax is None

UCAP fetcher doesn't accept float inputs for `n_participants`

Necessary for R interoperability

Push GitHub release to PyPI

See here and here

Add time-frequency analysis

Do time-frequency decomposition on unfiltered epochs with Morlet wavelets; export as single trial power (for LMMs; needs new dict of frequency bands of interest) + as evokeds (for CBPTs).

Users should be able to choose the divisive baseline mode for TFR

Creation of HTML reports fails if `ocular_correction=None`

Because no ica is created when skipping ocular correction